Twenty Product Images and a 30-Second Clip for One Listing: Made in the Browser on Replicate and fal, and What They Cost

A shop owner wants a listing dressed: twenty pictures of one product, in the ratios the marketplace wants, and a thirty-second clip for the top of the page. This page makes them in a browser on Replicate and fal, from one approved reference image, then shows the API route for when it is twenty listings rather than one, and only then adds up what the batch cost. Everything about the tools is read from their own documentation; nothing here rates one model's pictures against another's, because no host publishes such a figure.

What you hand over

Twenty PNG or JPG files of the product, on white, in use, in a setting, in 1:1 for the grid, 4:5 for the listing and 16:9 for the banner, all recognisably the same object. One 30-second MP4 built from three 10-second clips animated out of three of those pictures. And a note of the seeds and prompts that made the approved images, so the same look can be reproduced for the next product in the line.

Step one: the reference image

Do not start with twenty prompts. Start with one picture the client approves, either a real photograph of the product or a first generation you iterate until they say yes. Replicate's FLUX 1.1 pro form has an image_prompt input, described in its schema as an image used together with the text prompt to guide the generation toward the composition of that image; that is how the other nineteen stay on brand. fal's Wan 2.5 image-to-video takes an image as the first frame of the clip. One approved picture anchors everything else on this page.

Path A: the Playground tab

Replicate's docs say every model has its own playground page with a web form for running it in the browser, and that it is the place to start before touching the API. Open black-forest-labs/flux-1.1-pro, and the form has the fields from the schema: prompt; aspect_ratio from a list (1:1 by default, 16:9, 3:2, 2:3, 4:5, 5:4, 9:16, 3:4, 4:3, or custom with width and height between 256 and 1440 in multiples of 32); output_format, webp by default, or jpg or png; output_quality from 0 to 100, default 80; seed; safety_tolerance from 1 to 6, default 2; and prompt_upsampling, which rewrites your prompt for more creative results and is off by default. Set the format to png or jpg, because some shop uploaders will not take webp. Click Run, wait, download. Predictions made in the web interface are kept indefinitely on your dashboard, with a Delete button when you want them gone.

fal's docs describe the same shape: every model page has a Playground tab whose form is generated from the input schema, with required fields marked, an upload box for image inputs and a Run button; each result stays on screen so you can compare runs. The API tab shows the schemas, and the page shows the price and the average latency. If you have not chosen a model yet, fal's Sandbox runs one prompt through several models side by side and shows the estimated cost for the selected models and quantity before you submit.

Twenty images from one approved seed

The schema line that makes a series possible is seed: set it, the description says, for reproducible generation. When the client approves an image, write down its seed and prompt. Then produce the listing: the same product on a plain background, held in a hand, on a table, in the room it belongs in, each at the ratios the marketplace wants. Keep output_quality at 80 for the grid and raise it to 100 for the hero. Twenty runs of the form is twenty files; at the pace of the playground that is an afternoon, and it is the reason the API route below exists.

Thirty seconds is three clips

fal's Wan 2.5 image-to-video page sets the limits. It generates 5 or 10 seconds per run, at 480p, 720p or 1080p with 1080p as the default. The first-frame image must be JPEG, JPG, PNG without an alpha channel, BMP or WEBP, between 360 and 2000 pixels on each side, and at most 10MB. The motion prompt is up to 1500 characters, there is a negative prompt for what to avoid, and prompt expansion by a language model is on by default. Processing takes 1 to 3 minutes, longer for a 10-second clip. So a 30-second listing clip is three 10-second runs from three of your approved stills, joined in any video editor with a cut or a crossfade. The page carries a Commercial use badge; the licence section below says why to read it anyway.

Path B: the API, for the twentieth listing

The playground is where you find the prompt, the seed and the settings. The API is where you repeat them. fal's docs say every Playground result includes generated code that reproduces the exact request, in Python, JavaScript and cURL; the Python shape is:

import fal_client
result = fal_client.subscribe("fal-ai/nano-banana-2", arguments={"prompt": "a futuristic cityscape at sunset", "aspect_ratio": "16:9"})
print(result["images"][0]["url"])

Swap the model id and the arguments for the ones on your Wan or FLUX page. Replicate's docs show the equivalent, replicate.run("black-forest-labs/flux-schnell", input={"prompt": "A majestic lion"}), and say the client returns file objects you read and write to disk rather than bare URLs. The docs also explain cold boots: a model nobody has used for a while is switched off and can take minutes to start, but you are only charged for the time the prediction actually runs.

Save the files the same hour

One line in Replicate's data-retention docs decides your workflow. For predictions created through the API, all input parameters, output values, output files and logs are automatically removed after an hour by default; if you want them, save your own copies. Web-interface predictions are kept until you delete them. So the API script that makes two hundred images must also write them to disk as it goes, and the pictures in your dashboard from the playground are the only ones the host keeps for you.

A batch of twenty, and thirty seconds

The two charts above price this exact job at each published rate. Replicate prints per-image prices for the FLUX family: $0.04 on flux-1.1-pro, $0.025 on flux-dev, and $3.00 per thousand on flux-schnell, so twenty images are $0.80, $0.50 or $0.06. Video is per second everywhere: fal's Wan 2.5 at $0.05, Replicate's wan-2.1-i2v-480p at $0.09, OpenAI's sora-2 at $0.10 for 720p and sora-2-pro at $0.30 for 720p, $0.50 at 1024p and $0.70 at 1080p, with batch rates at half. Thirty seconds is $1.50, $2.70, $3.00 or $9.00. OpenAI prices its gpt-image models per million image tokens (gpt-image-1-mini at $8.00 per million output tokens) and points to a calculator instead of a per-image figure, so they are not on the image chart. The AI job cost calculator runs these for any batch size.

How each host bills

fal sells prepaid credits, drawn down as you use the platform, and charges only for successful outputs, never for server errors or queue time. Replicate bills most models by the time they run, with setup and idle time free on public models, and bills in arrears: the previous month's usage is charged at the start of the month, and the first time usage crosses certain limits an account can be charged early as a fraud check. Together AI is pay as you go with non-refundable fees; credits funded by ACH or wire are not usable until the payment clears, and access can be suspended on notice if any amount is 5 days past due. Together's GPU prices, $3.99 an hour for an H100 on demand or $1.99 preemptible, belong to the own-model page rather than here.

Who owns the mug's picture

Two licences sit on every generated image. The host's: fal says the customer owns and retains all right, title and interest in the Customer Input, subject to a licence for fal to process it, and keeps ownership of its own services and models; Together says you exclusively own Your Content and Output and grant it only a limited licence to host them. Then the model's: Together's terms note the models may carry their own third-party terms, and Replicate's FLUX 1.1 pro page says that by using it through Replicate you agree to the Black Forest Labs API agreement and terms of service. A picture going onto a product page needs both licences read, and fal's Commercial use badge on the Wan page is the host's word, not the model owner's.

What expires, and what can vanish

fal's credits expire 365 days from purchase, promotional credits in 90 days, and are non-refundable and non-transferable; an account that does not log in or use the services for more than a year can be deactivated. Replicate's terms say it may suspend or disable access to a Marketplace Model at any time for any reason, with no liability for the change. A catalogue built on one model has one point of failure, so test a second model on the same reference image while the first is still working, and keep its seeds too.

If the images sell a product

If the twenty pictures are for a print-on-demand mug rather than a stocked one, the image bill is the smallest line on the sheet. Printify's base price, shipping and the marketplace's cut are the numbers that decide the margin, and the POD pages on this site carry them. Put the eighty cents of images on the same sheet as the product.

What to do today

Pick one product you already sell. Get one reference image approved. Make five variations in the FLUX 1.1 pro playground with the seed written down, one 10-second Wan 2.5 clip from the best of them, and download everything. Time it. The host bill will be under a dollar; the hour you spent is the cost, and the seeds in your notes are what make the next product take half as long.